This retrospective study was conducted in accordance with the Declaration of Helsinki and was approved by the Ethics Committee of The First Affiliated Hospital of Zhengzhou University (Approval No 2023-KY-0223-002). Given the retrospective nature of this study, the requirement for informed consent was waived by the ethics committee. All patient data were anonymized and deidentified prior to analysis. Personal identifiers were removed, and each patient was assigned a unique study code to ensure confidentiality throughout the research process. No financial compensation was provided to the participants.

This retrospective study enrolled 160 treatment-naïve patients with ADG who underwent maximal safe surgical resection with histopathological confirmation at our institution between June 2019 and September 2021. The postoperative treatment regimen followed the standard protocol, consisting of concurrent chemoradiotherapy followed by adjuvant temozolomide chemotherapy. In cases of disease progression, targeted therapies such as bevacizumab or BRAF (B-Raf proto-oncogene, serine/threonine kinase) inhibitors (eg, for tumors harboring the BRAF V600E mutation) were administered based on clinical indications [19].

Inclusion criteria were as follows: (1) age ≥18 years, diagnosis of ADG in accordance with the 5th edition of the WHO Classification of Tumors of the Central Nervous System (World Health Organization [WHO] grade 3-4); (2) completion of preoperative magnetic resonance imaging or computed tomography evaluation; (3) no prior antitumor therapy; (4) receipt of standard treatment protocol; (5) availability of complete clinical data, including general characteristics, preoperative blood tests, cardiac, hepatic, and renal function assessments, as well as postoperative immunohistochemistry and molecular profiling; and (6) complete follow-up records.

Exclusion criteria were (1) postoperative mortality within 30 days due to surgical complications or nonneoplastic causes; (2) presence of severe systemic disease or active infection potentially affecting study outcomes; (3) diagnosis of other concurrent malignancies; (4) significant cardiac, hepatic, or renal dysfunction; and (5) history of prior cranial surgery.

Postoperative follow-up included clinical and radiological evaluations every 3 months during the first year and every 6 months thereafter, until October 1, 2024. Progression-free survival (PFS) was defined as the time from diagnosis to radiographic or clinical disease progression or death from any cause. OS was defined as the time from diagnosis to death or the last follow-up.

Clinical data were collected from patients with ADG who met the inclusion and exclusion criteria. The variables analyzed in this study included: clinical characteristics, molecular biomarkers, genetic alteration markers, preoperative peripheral hematological indicators, and inflammation-related ratios.

The study design is illustrated in Figure 1.

A random seed was set to divide the 160 cases into a training cohort (n=112) and a validation cohort (n=48) at a 7:3 ratio. The primary endpoint was OS, and the secondary endpoint was PFS. A total of 32 variables were included, encompassing clinical characteristics (age, gender, tumor location, etc), molecular biomarkers (Ki-67, IDH1, R132H, etc), and inflammatory indicators (NPAR, NLR, etc).

Continuous variables were dichotomized using the median as the cutoff. This approach was chosen to simplify the clinical interpretation of the model and to facilitate risk stratification, while also mitigating the impact of potential outliers and meeting the model’s assumption of proportional hazards for categorical predictors. Categorical variables were expressed as frequencies (percentages), and group comparisons were conducted using the chi-square test (expected frequency ≥5) or Fisher exact test (expected frequency <5). The Shapiro-Wilk test was used to assess the normality of continuous variables. For normally distributed data, comparisons between groups were performed using the 2-tailed independent-samples t test (2 groups) or 1-way ANOVA (3 or more groups), with data presented as mean (SD). For nonnormally distributed data, comparisons were performed using the Mann-Whitney U test (2 groups) or the Kruskal-Wallis H test (3 or more groups), with post hoc pairwise comparisons adjusted using the Bonferroni method. These data were reported as median (IQR). A 2-tailed P value <.05 was considered statistically significant for all analyses.

A total of 3 machine learning algorithms were applied: LASSO regression (using the glmnet package with 10-fold cross-validation to optimize λ), XGBoost (using the xgboost package with max depth=6 and learning rate=0.01), and RF (using the randomForest package with ntree=1000). Venn diagrams were used to identify overlapping important variables selected by all 3 algorithms. Survival differences between groups were visualized using Kaplan-Meier (KM) curves and compared using the log-rank test.

A Cox proportional hazards regression model was constructed to estimate hazard ratios (HRs) and 95% CIs. Based on multivariate Cox results, a nomogram was developed using the rms package. Time-dependent receiver operating characteristic (ROC) curves were generated, and the area under the curve (AUC) at 1, 3, and 5 years was calculated. The concordance index (C-index) was used to evaluate the model’s discrimination ability. Calibration was assessed using calibration curves based on 1000 bootstrap resamples. KM survival analysis based on the median nomogram score was performed to evaluate clinical utility. The validation cohort was used to assess the model’s predictive performance.

All statistical analyses were conducted using R software (version 4.4.1; R Foundation for Statistical Computing). A 2-sided P value <.05 was considered statistically significant.

Among the 160 patients included in this study, a total of 132 (82.5%) experienced death events, and 152 (95%) experienced disease progression. The median OS was 23.0 months (IQR 15.0-38.3 months), while the median PFS was 10.5 months (IQR 6.8-17.5 months). The 1-, 3-, and 5-year OS rates were 85.6%, 27.1%, and 12.8%, respectively. The corresponding 1-, 3-, and 5-year PFS rates were 40.0%, 6.5%, and 3.4%, respectively. KM survival curves are presented in Figure 2.

A total of 160 patients were included in this study, with a male-to-female ratio of approximately 1.4:1. The mean age was 51.3 (SD 13.6) years, and the median age was 53 (IQR 44-61) years. The maximum tumor diameter ranged from 1.0 to 9.5 cm, with a mean of 4.7 (SD 1.7) cm and a median of 5.0 cm (IQR 3.4-5.9 cm). Tumors were predominantly unilateral, with 76 cases (47.5%) on the left, 80 (50%) on the right, and 2 (2.5%) involving both sides. Midline shift was observed in 59.4% (95/160) of patients. All cases were classified as WHO grade 3-4. No significant differences in baseline characteristics were found between the training and validation cohorts (all P>.05). Detailed information is provided in Table 1.

Since the measurement data did not meet the assumption of normality according to the Shapiro-Wilk test, the nonparametric Wilcoxon rank-sum test was applied to evaluate the impact of different clinical characteristics on PFS and OS. Data are presented as median (IQR).

Using PFS as the outcome variable, significant differences (all P<.05) were observed in the following 10 variables: age (P<.001), tumor location (P=.03), NPAR (P=.002), NLR (P=.002), PNI (P=.01), SYN (P=.003), MGMT promoter methylation (P=.004), IDH1 R132 (P<.001), 1pLOH (P=.04), and 19qLOH (P<.001).

Using OS as the outcome variable, a total of 12 variables showed significant associations (all P<.05): age (P<.001), NPAR (P<.001), LMR (P=.03), NLR (P=.02), MLR (P=.03), PNI (P=.01), LDH (P=.04), SYN (P=.009), MGMT (P=.04), IDH1 R132 (P<.001), 1pLOH (P=.003), and 19qLOH (P<.001). Detailed results are presented in Table 2.

ADG, characterized by high malignancy and strong invasiveness, accounts for 45% of all gliomas [20,21]. The tumor heterogeneity and infiltrative growth patterns of ADG often lead to progressive neurological impairment, posing significant challenges for clinical management [9,22]. This study demonstrated that standardized comprehensive treatment (surgery combined with concurrent chemoradiotherapy and temozolomide) extended the median overall survival (mOS) to 23.0 (IQR 15.0-38.3) months, which is superior to the approximately 15 months reported in previous studies [2,23,24]. However, 95% (152/160) of patients experienced postoperative recurrence, with a median progression-free survival (mPFS) of only 10.8 months and a 5-year PFS rate of merely 1.9% (3/160). Moreover, the long-term prognosis remains poor, with a 5-year OS rate of just 12.8% (20/160) [25]. Therefore, it is of great significance to further explore the key prognostic factors affecting patients with ADG, establish novel prognostic models, and develop effective preventive strategies.

The key methodological advance of our work lies in the robust feature selection achieved by integrating 3 distinct machine learning algorithms (LASSO, XGBoost, and RF) [26]. This integrative approach circumvents the inherent pitfalls of relying on a single method, be it traditional Cox regression or an individual machine learning model [13,14]. Consequently, it ensures that the identified prognostic factors (age, NPAR, and PMR) are consistently salient across different modeling paradigms, thereby significantly strengthening the credibility of our results.

Compared to existing prognostic models for ADG, our model demonstrates superior predictive performance. While Kim et al [27] reported a C-index of 0.70 using a single machine learning approach, the OS model in this study achieved C-indices of 0.731 and 0.763 in the training and validation cohorts, respectively. Furthermore, the model maintained high time-dependent AUC values at multiple timepoints (AUCs of 0.836, 0.820, and 0.801 for 2-, 3-, and 4-year OS, respectively) along with favorable calibration. Importantly, our model innovatively incorporates readily accessible hematological biomarkers (NPAR and PMR). This provides a more cost-effective and clinically feasible prognostic assessment tool compared to traditional models that rely solely on clinical and molecular parameters [7,8].

Analysis of 32 variables showed that longer survival was observed in patients with canonical molecular markers such as IDH1 mutation and MGMT promoter methylation, implicating them as potential prognostic factors. IDH1 mutations play a critical role in suppressing tumor metabolism and proliferation [28,29]. Patients with IDH1-mutant astrocytoma included had an mOS of 40.0 (IQR 33.6–NA, NA is censored data—some patients remained alive at follow-up cutoff, so the upper quartile was not observed) months, markedly better than the 22.2 (IQR 14.4–34.7) months observed in IDH1 wild-type patients. Similarly, MGMT methylation enhances sensitivity of tumor cells to alkylating agents such as temozolomide by inhibiting MGMT transcription and reducing DNA repair capacity, thereby significantly prolonging both PFS and OS [29-31]. In this study, a 30-year-old patient with IDH-mutant astrocytoma and MGMT methylation achieved a survival period exceeding 64 months, demonstrating significant survival benefit [32].

Furthermore, by integrating multiple machine learning methods, the study systematically identified age, the inflammation-nutrition marker NPAR, and the platelet quantity and function marker PMR as independent prognostic factors for patients with ADG [16]. These variables demonstrated significant predictive power in forecasting the prognosis of patients with ADG [33,34].

Age was confirmed as a key prognostic factor in our model. Patients aged ≥53 years had a higher risk of disease progression and significantly shorter survival time compared with those aged <53 years. This finding aligns with previous research trends regarding the prognosis of older adults with ADG, which may be related to reduced treatment tolerance and alterations in tumor biological characteristics in older individuals. It suggests that clinicians should pay attention to age stratification and develop personalized treatment plans accordingly [28].

This study is the first to identify NPAR as an independent prognostic factor for ADG. Patients with high NPAR levels had significantly shorter mOS and mPFS, along with a markedly increased risk of disease recurrence and worse prognosis. As a novel indicator that comprehensively reflects systemic inflammation and nutritional status, the NPAR has increasingly been recognized for its prognostic value in various cancers such as pancreatic and urological cancers in recent years. The NPAR integrates both systemic inflammation (reflected by neutrophil percentage) and nutritional status (reflected by serum ALB), 2 key pathophysiological processes known to influence cancer prognosis. Neutrophils can impair immune surveillance by suppressing natural killer cell function and promote tumor cell proliferation and metastasis [12,35]. Serum ALB levels not only reflect the body’s protein metabolic status but are also closely associated with immune dysfunction [36]. The prognostic value of NPAR may stem from its correlation with systemic inflammatory responses mediated by proinflammatory cytokines in the TME. This pathophysiological cascade can accelerate the progression of cachexia and facilitate the evolution of malignant tumor phenotypes by modulating the TME [37,38].

The prognostic value of NPAR stems from its integration of 2 pivotal pathophysiological pathways: protumor inflammation and host nutritional status. The neutrophil component reflects systemic inflammation. Within the glioma TME, neutrophils are polarized into tumor-associated neutrophils (TANs), which facilitate tumor progression by secreting matrix metalloproteinases to enhance invasion, releasing proangiogenic factors like vascular endothelial growth factor, and contributing to T-cell dysfunction [39]. Emerging evidence specifically links this protumorigenic neutrophil activity to glioma progression [40]. Conversely, hypoalbuminemia, indicative of poor nutritional status and diminished physiological reserve, is a well-established marker of impaired immune competence and poorer outcomes in patients with cancer [41]. Thus, NPAR comprehensively captures this detrimental interplay between a proinflammatory TME and a compromised host state.

Furthermore, this study identified a higher PMR as an independent prognostic factor for prolonged overall survival in patients with ADG. This finding is consistent with a study on angiosarcoma, which also reported that a high mean platelet volume–to–platelet count ratio served as an independent risk factor for patient mortality [42]. This consistent pattern across malignancies highlights the biological significance of composite indices based on platelet count and volume.

While studies have established that platelets can facilitate glioma progression by releasing factors like platelet-derived growth factor to stimulate proliferation [43] and by modulating the immune microenvironment to promote an M2 phenotype and immune escape [44], our discovery of PMR as a protective factor suggests a more nuanced role. We hypothesize that an elevated PMR reflects a distinct platelet phenotype characterized by a lower mean platelet volume which is generally associated with diminished platelet reactivity. Consequently, a high PMR may indicate a reduced capacity of platelets to engage in these documented protumorigenic processes, thereby explaining its association with improved survival in ADG.

The prognostic nomogram developed from the aforementioned factors demonstrates significant translational potential. This tool can be integrated into electronic medical record systems to enable automated risk stratification at the time of diagnosis. Consequently, high-risk patients could be considered for more aggressive treatment strategies—such as intensified chemoradiation following maximal safe resection—or more intensive follow-up (eg, neuroimaging every 2-3 months), and even enrollment in novel clinical trials. Conversely, low-risk patients may be suitable candidates for treatment de-escalation, such as shortened adjuvant chemotherapy cycles or reduced radiation doses, aiming to preserve oncological control while enhancing quality of life. This nomogram-based decision-making model represents a constructive step toward precision neuro-oncology.

This study is subject to several limitations, including its single-center retrospective design, which may limit generalizability, and a variable-to-sample size ratio that raises concerns about overfitting despite internal validation. These limitations underscore the necessity for external validation in independent multicenter cohorts as an essential next step. Moreover, the molecular mechanisms linking NPAR and PMR to glioma prognosis are not fully understood. To address these points and advance this research, future work must focus on multicenter prospective validation, augmenting the model’s power through the integration of radiomics and genomics, and leveraging single-cell and multi-omics technologies to decipher the functional roles of these circulating biomarkers in glioma pathogenesis.

In conclusion, this study significantly improved the robustness of prognostic biomarkers through a multimodel cross-validation strategy. Among the numerous factors, age, the preoperative inflammation-nutrition composite indicator NPAR, and platelet quantity and function marker PMR were key prognostic factors. Patients with ADG with younger age and lower NPAR values had a lower risk of disease progression and better prognosis; meanwhile, patients with higher PMR exhibited longer survival. These biomarkers are not only easily accessible and cost-effective but also highly beneficial, offering a new perspective for personalized treatment of ADG.